Rotated FCOS vs Oriented R-CNN on macOS

In June we ran Oriented R-CNN on a MacBook with --device mps — no CUDA toolchain, one CLI command, rotated boxes on a real aerial tile. v0.2.0 added a fourth detector family: Rotated FCOS, an anchor-free single-stage model with a decoded-rIoU 1× Hub checkpoint.

This post puts both on the same Mac and the same images. Same canvas, same NMS — Apple M1 Max, PyTorch MPS, Hub 1× weights. We also pull the training wall times from the published Hub runs (same NVIDIA L4 recipe). The question is practical: does the new one-stage model feel good enough on a laptop to replace the two-stage demo default?

Short answer: FCOS is the faster one-stage / cheaper-train / MPS demo. It is not more accurate: official Task 1 is 73.07% versus Oriented R-CNN 76.73%. Scores are less peaked, so copying a two-stage threshold onto FCOS will silently drop half the boxes. Use 0.20 for FCOS and 0.55 for Oriented R-CNN.

What we compare Link to heading

Both are ResNet-50 + FPN, DOTA le90, train+val pretrain. Published scores are official Task 1, not local val.

Inference numbers below: Apple M1 Max, 64 GB unified memory, PyTorch 2.13.0, oriented-det 0.2.0, --device mps. Protocol: deploy score floors, merge NMS IoU ≤ 0.1. Training wall times come from the Hub run logs on a single NVIDIA L4 (see below).

Quick start Link to heading

From an oriented-det checkout with the usual macOS install (uv pip install torch torchvision then uv pip install -e .):

odet pretrained download rotated_fcos_dota_le90_1x
odet pretrained download oriented_rcnn_dota_le90_1x

odet image-demo demo/demo.jpg hf://rotated_fcos_dota_le90_1x \
  --out-file demo_fcos.png \
  --device mps \
  --score-thr 0.20 \
  --nms-thr 0.1

Swap the slug for Oriented R-CNN and use --score-thr 0.55. Keep --nms-thr 0.1.

Bus lot: same scene as the June demo Link to heading

demo/demo.jpg is the 1024×1024 DOTA tile from the macOS walkthrough — diagonal buses and trucks, the scene where axis-aligned boxes look silly.

Box geometry looks right on both: headings follow the chevron parking, large-vehicle / small-vehicle labels match. The visible difference is score calibration — Oriented R-CNN piles many boxes near 1.00; FCOS spreads them across a wider band. That is expected for a one-stage sigmoid head versus a two-stage ROI classifier, but it changes which CLI threshold you want.

Score thresholds: do not copy a two-stage --score-thr onto FCOS Link to heading

The June post uses --score-thr 0.55 for Oriented R-CNN (the Hub deploy floor). A stricter overlay such as 0.7 still works on that family. On FCOS it does not: a two-stage threshold hides most of the scene.

For FCOS demos, start at 0.20. For Oriented R-CNN, 0.55. Keep NMS at 0.1.

Latency on MPS Link to heading

Timed after warmup; mean of 5 single-forward runs (or 3 for the tiled harbor). Canvas 1024×1024, ORIENTED_DET auto window batch 32 on MPS.

On sparse tiles the one-stage head is almost 2× faster. On the dense bus lot the gap shrinks because final oriented NMS (Python, AABB-prefiltered) scales with detection count. The harbor tile (six overlapping windows) still favors FCOS by about 40%.

Neither model needs custom CUDA kernels. MPS just works.

Training wall time (same L4 recipe) Link to heading

The Hub 1× checkpoints were trained on the same DOTA tile recipe: 13,691 train+val tiles, batch size 2, 12 epochs, single NVIDIA L4. Timings are from the published sidecar logs (oriented_rcnn_r50_fpn_dota_le90_1x-725c244f.log, rotated_fcos_r50_fpn_dota_le90_1x-a87b6dba.log).

Rotated FCOS finishes the 1× schedule in an afternoon — about 4.5× less wall clock than Oriented R-CNN on the same GPU and data.

That gap is architectural: Oriented R-CNN pays for an RPN plus oriented RoIAlign on every proposal every step. FCOS is a single dense head over P3–P7. The July ProbIoU post already showed Rotated Faster R-CNN (~58 min/epoch, 11h 30m 1× wall) beating Oriented R-CNN on training cost; FCOS lands in a similar per-epoch band while staying one-stage and anchor-free.

Harbor: sliding-window ships Link to heading

Same large.jpg as the v0.1.1 harbor demo (1299×1904). The DOTA recipe tiles oversized rasters; here that is six 1024 windows with 200 px overlap.

odet image-demo demo/large.jpg hf://rotated_fcos_dota_le90_1x \
  --out-file large_fcos.png \
  --device mps \
  --score-thr 0.20 \
  --nms-thr 0.1

Visually both cover the moored rows; FCOS finishes in under two seconds on this machine.

Planes and storage tanks Link to heading

Two more DOTA val tiles for class variety — airport apron and tank farm — same deploy thresholds.

Both find the aircraft. FCOS scores sit lower; Oriented R-CNN saturates near 1.00. Parking-lot vehicles are comparable.

On official Task 1, FCOS is not ahead of Faster R-CNN on tanks (84.28 vs 84.41). On this tile both FCOS and Oriented R-CNN land the visible tanks cleanly.

When to pick which Link to heading

For laptop demos after v0.2, Rotated FCOS 1× is the better speed default than Oriented R-CNN: fewer parameters, faster MPS inference, cheaper 1× training, and no RPN. Oriented R-CNN still wins Task 1. Keep FCOS at 0.20 and Oriented R-CNN at 0.55, NMS 0.1.

Commands (copy-paste) Link to heading

# Prefetch
odet pretrained download rotated_fcos_dota_le90_1x
odet pretrained download oriented_rcnn_dota_le90_1x

# FCOS on the June bus-lot tile
odet image-demo demo/demo.jpg hf://rotated_fcos_dota_le90_1x \
  --out-file demo_fcos.png --device mps --score-thr 0.20 --nms-thr 0.1

# Side-by-side Oriented R-CNN
odet image-demo demo/demo.jpg hf://oriented_rcnn_dota_le90_1x \
  --out-file demo_orcnn.png --device mps --score-thr 0.55 --nms-thr 0.1

# Harbor (sliding windows)
odet image-demo demo/large.jpg hf://rotated_fcos_dota_le90_1x \
  --out-file large_fcos.png --device mps --score-thr 0.20 --nms-thr 0.1

References Link to heading

• oriented-det on GitHub
• Pretrained zoo — rotated_fcos_dota_le90_1x, oriented_rcnn_dota_le90_1x (sidecar .log files hold the train timing summaries)
• Rotated FCOS recipes
• Previous: Oriented-Det v0.2.0 · macOS Oriented R-CNN walkthrough · Faster R-CNN training-cost context
• Next: A static demo of three oriented detectors

Written on September 2, 2026 by Jeff Faudi. Link to heading